Research On Deep Intelligent Fault Diagnosis Method Of Rolling Bearings With Few Samples

Rolling bearings are one of the core components of various mechanical equipment.Its high reliability is the basic guarantee for the normal operation of most mechanical equipment.Therefore,real-time health monitoring of rolling bearings is a strong guarantee for the normal operation of mechanical equipment.Traditional rolling bearing fault diagnosis relies heavily on diagnostic experts.However,relying on expert fault diagnosis methods are time-consuming and laborious.The existing fault diagnosis based on deep learning is too dependent on a large amount of data training model,but in practical engineering,a large amount of data can’t be collected for various reasons,so the method based on deep learning is also difficult to get a good application in the actual rolling bearing fault diagnosis.In view of the above problems,this thesis studies the deep intelligent fault diagnosis method of rolling bearings under few samples based on deep learning.The main research contents are as follows:(1)A rolling bearing fault test bench is built and the running state of the rolling bearing is collected in real time.The whole process includes bearing fault processing,test setting and data acquisition.Then the collected data is standardized and divided into samples.Finally,the data is constructed into a few-shot learning task.(2)A few-shot fault diagnosis method based on Earth Mover ’s Distance(EMD)and multiattention is proposed.This method decomposes the features into local descriptors,which makes full use of all the information contained in the fault samples,and uses EMD distance to measure the distance between local descriptors.A cross-reference mechanism is introduced to assign different weights to different descriptors,which solves the problem of insufficient utilization of rolling bearing descriptors under few samples.In addition,in order to solve the problem of low efficiency of feature extraction,an attention module is introduced into the feature extraction network to improve the efficiency of feature extraction.Finally,the effectiveness of the method under few samples is verified by the rolling bearing fault diagnosis experiment task.(3)A fault diagnosis method of rolling bearing under few samples based on convex optimization meta-learning is proposed.This method uses a support vector machine(SVM)linear classifier as the base learner of meta-learning.Using the convex nature of the objective function of the SVM classifier,the meta-learning parameter optimization is converted into a convex optimization,which can be easily solved by quadratic programming.The optimal solution can be converted into a dual problem.The number of optimized parameters in the dual formula is much smaller than the feature dimension,so it also solves the over-fitting problem caused by the increase of data dimension.This method uses the advantages of meta-learning to improve the problem of insufficient learning ability based on metric learning,and improves the accuracy of rolling bearing fault diagnosis under strong noise.The proposed method is applied to rolling bearing fault diagnosis experiments to verify the generalization performance of the method under few samples and strong noise.